What's happening
Anthropic's Claude Code is driving adoption of "agentic software engineering" - autonomous AI agents that plan, execute, and debug code through terminal interfaces. Built on Claude 4.5, the tool scored 77.2% on SWE-bench Verified, ahead of GPT-5.2's 74.1%.
The shift: engineers are moving from conversational coding to structured agent orchestration. High-performance users rely on Plan Mode - generating markdown specifications for review before execution. This "review-first" pattern is becoming standard practice.
The technical edge
Claude Code's 200k-token context window (roughly 150k words) exceeds GitHub Copilot's ~8k and GPT-4 Turbo's 128k. This capacity enables entire codebase analysis and persistent project memory across sessions.
Key enterprise features:
- CLAUDE.md manifests for defining tech stacks and linting rules per repository
- OAuth-based directory scoping for compliance
- VS Code integration with multi-file change tracking
- Sub-agent delegation for parallel task execution
Pricing sits at $100-$200 monthly through Claude Max subscriptions, with enterprise tiers supporting multi-agent orchestration.
What's working in practice
Matt Pocock's engineering workflows demonstrate asynchronous delegation - leaving agents to solve complex bugs overnight while maintaining system integrity through structured planning.
The community has standardized around CLAUDE.md files in repository roots, acting as system memory so project constraints don't require repetition. Power users are running in "dangerous mode" (skipping permission prompts) and relying on Git history as their safety net.
The limitations
This augments developers, it doesn't replace them. Early "vibe coding" without oversight risked structure loss. Users must review outputs to prevent drift - the tool requires human engagement for complex architectural decisions.
Security researcher Ethan Mollick flags risks in granting AI full file access, particularly in corporate environments.
What to watch
The pattern emerging: treat CLI agents as compilers of intent, not conversational partners. Success correlates with structured planning phases and architectural review before execution.
For CTOs evaluating agentic tools, the question isn't whether AI can write code. It's whether your team has the discipline to review plans before agents ship changes.